Embedded Files
EchoWrist (CHI'24)
A low-power wristband using active acoustic sensing to estimate 3D hand poses and recognize hand-object interactions with 97.6% accuracy, enabling efficient HCI applications.
SonicID (IMWUT'24)
A low-power authentication system for smart glasses, using ultrasonic face scanning and ResNet-18 to achieve 96.6% accuracy in authenticating users with minimal training data.
ActSonic (IMWUT'24)
Low-power smartglasses equipped with ultrasonic sensors that recognize 27 everyday activities, like eating or brushing teeth, with up to 93.4% accuracy in real-world home settings.
Grab-n-Go
Under Submission.
Details will be updated upon acceptance to venue.
Open-Vocabulary Segmentation for Remote Sensing
Open-Vocabulary Segmentation for Remote Sensing
Developing a model for satellite imagery that enables zero-shot segmentation using natural language descriptions of unseen classes by integrating GRAFT, a foundation model for aligning satellite images and text, with SegViT and CLIPSeg architectures to address challenges in remote sensing datasets.
Low-Resolution Change Event Discovery for Heritage Sites
Low-Resolution Change Event Discovery for Heritage Sites
This project focuses on improving satellite-based change detection models with an emphasis on preserving heritage sites. By combining low-resolution imagery for initial detection with high-resolution data for detailed analysis, we aim to provide a cost-efficient framework for scalable and accurate monitoring.
Towards Forgetting and Online Machine Unlearning
Towards Forgetting and Online Machine Unlearning
This project focuses on developing methods for online unlearning, enabling the removal of data influence from models without requiring retraining. By introducing novel algorithms, such as Online Gradient Descent Forgetting and a tree-based approach for multiple deletions, it addresses privacy challenges in dynamic, real-time learning environments.
Page updated
Google Sites
Report abuse